Documentation / Start

Architecture

How evidence, interpretation, review and application contracts fit together.

One package inside your service

Your application owns its HTTP interface, authentication and database resources. Ontave runs in the server process and stores its internal tables under the ontave PostgreSQL schema.

text
Authenticated host context
          ↓
     Ontave runtime
          ├─ Evidence + durable jobs → PostgreSQL
          ├─ Scoped worker → explicit model adapter
          ├─ Staged bundles → authorized review
          ├─ Candidate → evaluation → activation
          └─ Pinned contract → accepted-data query

Separate records, separate meaning

Evidence preserves the source. Runs record interpretations and available provider usage. Bundles capture the content and dependencies a reviewer sees. Assertions retain accepted facts and their history. Projections support current structured reads.

Ontology versions describe the active logical model. Candidates represent proposed additive changes. Contracts describe the subset a particular application adopts. A contract is not a historical data snapshot.

Inference outside the transaction

A worker claims a fenced lease in a short database transaction. Model networking happens outside it. Staging uses another transaction that checks the lease, fence and model version. A provider can repeat a request after a crash; the guarantee concerns committed effects, not exactly-once inference.

Explicit boundaries

A receipt is not extraction completion. A candidate is not an active model. Model activation is not fact acceptance. Reviewer acceptance is not independent verification. These boundaries are represented by distinct operations and states.

Ontave Zero · Apache 2.0 · Alpha release status
Search concepts, guides and API methods.